List your company
ResearchFormerly WizWand

Mathematical Reasoning on AMC 23 (Acc, P@N, Tok, CR)

95Accuracy4B-Instruct
1.425.75074.3Feb 13, 2026Feb 19, 2026Feb 25, 2026Mar 4, 2026Mar 10, 2026Mar 16, 2026Mar 23, 2026
Updated 7mo ago

Evaluation results

MethodLinks
4B-InstructModel=4B-Instruct, RL Algorithm=HACPO, Heterogeneity Setting=Qwen3-4B + Qwen3-4B-Instruct2026.0395———
CoDBackbone=DeepSeek-R1-Distill-Qwen-7B, Training Paradigm=training-free2026.0290.5955,01874.6
DEERBackbone=DeepSeek-R1-Distill-Qwen-7B, Training Paradigm=training-free2026.0290.5954,86672.3
O1-PrunerBackbone=DeepSeek-R1-Distill-Qwen-7B, Training Paradigm=training-free2026.0290.5955,37980
4B-InstructModel=4B-Instruct, RL Algorithm=GSPO, Heterogeneity Setting=Qwen3-4B + Qwen3-4B-Instruct2026.0390———
4B-InstructModel=4B-Instruct, RL Algorithm=GSPO×2, Heterogeneity Setting=Qwen3-4B + Qwen3-4B-Instruct2026.0390———
COTBackbone=DeepSeek-R1-Distill-Qwen-7B, Training Paradigm=training-free2026.0288.5956,722100
CRSTBackbone=DeepSeek-R1-Distill-Qwen-7B, Training Paradigm=SFT-based2026.0288.592.56,33794.2
LASERBackbone=DeepSeek-R1-Distill-Qwen-7B, Training Paradigm=training-free2026.0288.595.53,13746.6
ThinkPruneBackbone=DeepSeek-R1-Distill-Qwen-7B, Training Paradigm=training-free2026.0288953,39750.5
ER-RLBackbone=DeepSeek-R1-Distill-Qwen-7B, Training Paradigm=RL-based2026.0287.5954,22162.8
On-Policy SFTBackbone=DeepSeek-R1-Distill-Qwen-7B, Training Paradigm=SFT-based2026.0287.597.51,96129.2
4B-InstructModel=4B-Instruct, RL Algorithm=GRPO, Heterogeneity Setting=Qwen3-4B + Qwen3-4B-Instruct2026.0387.5———
4BModel=4B, RL Algorithm=HACPO, Heterogeneity Setting=Qwen3-4B + Qwen3-4B-Instruct2026.0385———
4B-InstructModel=4B-Instruct, RL Algorithm=Base, Heterogeneity Setting=Qwen3-4B + Qwen3-4B-Instruct2026.0385———
L1Backbone=DeepSeek-R1-Distill-Qwen-7B, Training Paradigm=training-free2026.0283.592.52,19132.6
4BModel=4B, RL Algorithm=GRPO, Heterogeneity Setting=Qwen3-4B + Qwen3-4B-Instruct2026.0377.5———
L1Backbone=DeepSeek-R1-Distill-Qwen-1.5B, Training Paradigm=training-free2026.0274.5902,17023.7
On-Policy SFTBackbone=DeepSeek-R1-Distill-Qwen-1.5B, Training Paradigm=SFT-based2026.0273.5901,97721.6
COTBackbone=DeepSeek-R1-Distill-Qwen-1.5B, Training Paradigm=training-free2026.027392.59,171100
O1-PrunerBackbone=DeepSeek-R1-Distill-Qwen-1.5B, Training Paradigm=training-free2026.0272.5903,94242.9
TokenSkipBackbone=DeepSeek-R1-Distill-Qwen-7B, Training Paradigm=training-free2026.0272.572.59,459140.7
4B-InstructModel=4B-Instruct, RL Algorithm=Naive, Heterogeneity Setting=Qwen3-4B + Qwen3-4B-Instruct2026.0372.5———
ER-RLBackbone=DeepSeek-R1-Distill-Qwen-1.5B, Training Paradigm=RL-based2026.0272.492.54,91353.6
ThinkPruneBackbone=DeepSeek-R1-Distill-Qwen-1.5B, Training Paradigm=training-free2026.0272952,67229.1
LASERBackbone=DeepSeek-R1-Distill-Qwen-1.5B, Training Paradigm=training-free2026.0272904,26246.4
CoDBackbone=DeepSeek-R1-Distill-Qwen-1.5B, Training Paradigm=training-free2026.0271.587.58,51892.8
DEERBackbone=DeepSeek-R1-Distill-Qwen-1.5B, Training Paradigm=training-free2026.027192.56,90875.3
CRSTBackbone=DeepSeek-R1-Distill-Qwen-1.5B, Training Paradigm=SFT-based2026.0270.592.57,41080.8
4BModel=4B, RL Algorithm=GSPO, Heterogeneity Setting=Qwen3-4B + Qwen3-4B-Instruct2026.0367.5———
4BModel=4B, RL Algorithm=GSPO×2, Heterogeneity Setting=Qwen3-4B + Qwen3-4B-Instruct2026.0367.5———
4BModel=4B, RL Algorithm=Base, Heterogeneity Setting=Qwen3-4B + Qwen3-4B-Instruct2026.0365———
HACPOModel Backbone=Qwen3-8B-Base2026.0362.5———
HACPOModel Backbone=Qwen3-4B-Base2026.0360———
4BModel=4B, RL Algorithm=Naive, Heterogeneity Setting=Qwen3-4B + Qwen3-4B-Instruct2026.0360———
qwen3-4BModel=qwen3-4B, RL Algorithm=HACPO, Heterogeneity Setting=Qwen3-4B-Base + Llama3.2-3B-Instruct2026.0360———
ARE-GroupAModel=Qwen2.5 (7B-Instruct)2026.0360———
ARE-GroupBModel=Qwen2.5 (7B-Instruct)2026.0357.7———
4B-BaseModel=4B-Base, RL Algorithm=HACPO, Heterogeneity Setting=Qwen3-1.7B-Base + Qwen3-4B-Base2026.0357.5———
4B-BaseModel=4B-Base, RL Algorithm=GSPO×2, Heterogeneity Setting=Qwen3-1.7B-Base + Qwen3-4B-Base2026.0355———
qwen3-4BModel=qwen3-4B, RL Algorithm=GSPO×2, Heterogeneity Setting=Qwen3-4B-Base + Llama3.2-3B-Instruct2026.0355———
qwen3-4BModel=qwen3-4B, RL Algorithm=Naive, Heterogeneity Setting=Qwen3-4B-Base + Llama3.2-3B-Instruct2026.0355———
4B-BaseModel=4B-Base, RL Algorithm=GSPO, Heterogeneity Setting=Qwen3-1.7B-Base + Qwen3-4B-Base2026.0352.5———
qwen3-4BModel=qwen3-4B, RL Algorithm=GSPO, Heterogeneity Setting=Qwen3-4B-Base + Llama3.2-3B-Instruct2026.0352.5———
baseline-GroupBModel=Qwen2.5 (7B-Instruct)2026.0352.5———
baseline-GroupAModel=Qwen2.5 (7B-Instruct)2026.0350———
StepEntropyBackbone=DeepSeek-R1-Distill-Qwen-7B, Training Paradigm=RL-based2026.0248.577.59,734144.8
StepEntropyBackbone=DeepSeek-R1-Distill-Qwen-1.5B, Training Paradigm=RL-based2026.0247.577.59,237100.7
4B-BaseModel=4B-Base, RL Algorithm=GRPO, Heterogeneity Setting=Qwen3-1.7B-Base + Qwen3-4B-Base2026.0347.5———
4B-BaseModel=4B-Base, RL Algorithm=Naive, Heterogeneity Setting=Qwen3-1.7B-Base + Qwen3-4B-Base2026.0347.5———
qwen3-4BModel=qwen3-4B, RL Algorithm=GRPO, Heterogeneity Setting=Qwen3-4B-Base + Llama3.2-3B-Instruct2026.0347.5———
HACPOModel Backbone=Qwen3-1.7B-Base2026.0345———
1.7B-BaseModel=1.7B-Base, RL Algorithm=GSPO, Heterogeneity Setting=Qwen3-1.7B-Base + Qwen3-4B-Base2026.0345———
1.7B-BaseModel=1.7B-Base, RL Algorithm=HACPO, Heterogeneity Setting=Qwen3-1.7B-Base + Qwen3-4B-Base2026.0345———
4B-BaseModel=4B-Base, RL Algorithm=Base, Heterogeneity Setting=Qwen3-1.7B-Base + Qwen3-4B-Base2026.0340———
qwen3-4BModel=qwen3-4B, RL Algorithm=Base, Heterogeneity Setting=Qwen3-4B-Base + Llama3.2-3B-Instruct2026.0340———
1.7B-BaseModel=1.7B-Base, RL Algorithm=GRPO, Heterogeneity Setting=Qwen3-1.7B-Base + Qwen3-4B-Base2026.0337.5———
1.7B-BaseModel=1.7B-Base, RL Algorithm=GSPO×2, Heterogeneity Setting=Qwen3-1.7B-Base + Qwen3-4B-Base2026.0337.5———
Base ModelModel=Qwen2.5 (7B-Instruct)2026.0337.5———
llama3.2-3BModel=llama3.2-3B, RL Algorithm=HACPO, Heterogeneity Setting=Qwen3-4B-Base + Llama3.2-3B-Instruct2026.0335———
TokenSkipBackbone=DeepSeek-R1-Distill-Qwen-1.5B, Training Paradigm=training-free2026.0232.532.517,452190.3
1.7B-BaseModel=1.7B-Base, RL Algorithm=Naive, Heterogeneity Setting=Qwen3-1.7B-Base + Qwen3-4B-Base2026.0332.5———
ARE-GroupAModel=Qwen2.5 (1.5B-Instruct)2026.0332.5———
1.7B-BaseModel=1.7B-Base, RL Algorithm=Base, Heterogeneity Setting=Qwen3-1.7B-Base + Qwen3-4B-Base2026.0330———
Base ModelModel=Qwen2.5 (1.5B-Instruct)2026.0330———
HACPOModel Backbone=Llama3.2-3B-Instruct2026.0327.5———
ARE-GroupCModel=Qwen2.5 (1.5B-Instruct)2026.0327.5———
ARE-GroupCModel=Llama3.1 (8B-Instruct)2026.0327.5———
llama3.2-3BModel=llama3.2-3B, RL Algorithm=GRPO, Heterogeneity Setting=Qwen3-4B-Base + Llama3.2-3B-Instruct2026.0325———
llama3.2-3BModel=llama3.2-3B, RL Algorithm=GSPO, Heterogeneity Setting=Qwen3-4B-Base + Llama3.2-3B-Instruct2026.0322.5———
llama3.2-3BModel=llama3.2-3B, RL Algorithm=Naive, Heterogeneity Setting=Qwen3-4B-Base + Llama3.2-3B-Instruct2026.0322.5———
baseline-GroupAModel=Qwen2.5 (1.5B-Instruct)2026.0322.5———
baseline-GroupCModel=Qwen2.5 (1.5B-Instruct)2026.0322.5———
baseline-GroupCModel=Llama3.1 (8B-Instruct)2026.0322.5———
ARE-GroupBModel=Llama3.1 (8B-Instruct)2026.0322.5———
HACPOModel Backbone=Llama3.2-1B-Instruct2026.0320———
llama3.2-3BModel=llama3.2-3B, RL Algorithm=Base, Heterogeneity Setting=Qwen3-4B-Base + Llama3.2-3B-Instruct2026.0320———
baseline-GroupBModel=Llama3.1 (8B-Instruct)2026.0320———
llama3.2-3BModel=llama3.2-3B, RL Algorithm=GSPO×2, Heterogeneity Setting=Qwen3-4B-Base + Llama3.2-3B-Instruct2026.0317.5———
HACPOModel Backbone=Llama3.2-1B-Instruct, Experimental Context=Heterogeneous Setup2026.0315———
Base ModelModel=Llama3.1 (8B-Instruct)2026.035———